Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: Modeling and simulation of atomization is challenging due to the existence of a wide range of length
Received 13 January 2015 scales. This multiscale nature of atomization introduces a fundamental challenge to numerical simula-
Received in revised form 1 June 2015 tion. A pathway to comprehensive modeling is still to be found. The present study proposes a multiscale
Accepted 9 July 2015
multiphase flow model for atomization simulations, where the large-scale interfaces are resolved by the
Available online 17 July 2015
Volume-of-Fluid (VOF) method and the small droplets by the Lagrangian point-particle (LPP) model.
Particular attention is focused on the momentum coupling between LPP and resolved flow and the con-
Keywords:
version between droplets represented by VOF and LPP. A series of multiphase flow problems are consid-
Multiphase flows
Atomization
ered to validate the model. The results obtained by a number of simulations are compared against direct
Lagrangian point-particle numerical simulation (DNS) results and experimental data. In particular, the model is applied to simulate
Volume of fluid method the gas-assisted atomization experiment, and the numerical results are compared to the experimental
Multiscale modeling measurements for a quantitative validation.
Ó 2015 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ijmultiphaseflow.2015.07.002
0301-9322/Ó 2015 Elsevier Ltd. All rights reserved.
Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 123
and in shear layers. Nevertheless, even with AMR, DNS is still often are evolved in a Lagrangian framework. In atomization, the typical
too expensive for many atomization problems. First of all, it is well Weber number of the droplets, Wep , is usually very small. (Precise
known that octtree AMR adds a large overhead to computations. definition of Wep will be given in Section ‘Modeling and numerical
This is mainly due to the placement of objects in memory, resulting methods’.) As a result, the droplets remain close to a spherical
numerous page faults. Because 3D block-AMR requires the creation shape, and the LPP model is suitable to capture their motion and
of a very large number of boxes, similar difficulties may also apply the backward effect to the flow. From the filtering perspective,
in that case. Second, the parallelization of AMR adds additional LPP can be viewed as a local spatial filter with a variable filtering
challenges. Regions in which a high density of small computations length scale (Climent and Magnaudet, 2006). Since it is difficult
cells exist are not always distributed uniformly in space (in to filter the overall flow field, filtering is only applied at the particle
atomization they are close to the nozzle) which makes domain location with a filtering length scale which is related to the droplet
decomposition into blocks of equal size difficult. Third, the most size. In this work, a multiscale simulation approach that couples
demanding mesh requirement is usually introduced by the small the LPP model and an interface-capturing scheme is proposed. In
droplets that are formed in atomization. Typically, to obtain rea- this approach, the goal is to filter the scales on the order of a small
sonable results of motion and interface dynamics for a spherical number, say nc , of grid cells, and to represent larger scales. Thus
droplet, at least eight to twelve grid points across the droplet droplets of a diameter smaller than or of the order of nc Dx are rep-
diameter are required. It is known from experiments that the resented by the LPP model, while larger droplets and fluid masses
diameter of the atomizing liquid jet is about three orders of mag- are represented by an interface tracking scheme. Thus part of the
nitude larger than the smallest droplets. For such a wide range of flow is represented in a DNS-like manner while both the flow
length scales, the number of grid points that is required to resolve inside the small droplets and the surrounding flow perturbed by
every scale, namely a true DNS, is extremely expensive even with these droplets are filtered. Thus the motion of the small droplets
AMR. and their backward effect to the resolved flow are represented by
An alternative approach to resolve this multiscale challenge in the LPP force model. Moreover, care is taken to work with
atomization simulation is to introduce subgrid models, although Reynolds and Weber numbers sufficiently moderate to minimize
the two ideas could be combined. The idea of subgrid modeling the energy present at such small scales.
is very simple: only the physical scales larger than the cell size In recent years, several pioneering efforts have been made to
are resolved, while for the physical scales that are smaller suitable couple the LPP model with interface-capturing schemes, such as
models are introduced to represent the relevant physics. A good LS (Herrmann, 2010) and VOF (Tomar et al., 2010). The results
example of subgrid modeling is the Large-Eddy Simulation (LES) are quite encouraging and show the great potential of this multi-
model for turbulence (Pope, 2000). In LES, a low-pass filtering is scale approach for accurate large-scale simulations of atomization.
applied and only the large eddies are resolved. As a result, a coarser However, there still exist unresolved issues in the coupling of LPP
mesh compared to DNS of turbulence can be used. However, to with interface-capturing schemes which require further investiga-
accurately capture the large-scale physics, the small unresolved tion and validation. In conventional simulations with the LPP
turbulent eddies and their interaction with the large scales have model the grid cell is much larger than the particle size, typically
to be represented by a subgrid model, such as the Smagorinsky dp < Dx=10, where dp and Dx represent the particle diameter and
model (Smagorinsky, 1963). For turbulent multiphase flows like cell size, respectively, as shown in Fig. 1(a). As a result, a single
atomization, subgrid modeling is more challenging since both the computational cell may contain many particles. Sometimes, the
unresolved turbulence and multiphase scales, and the interaction number of particles in a cell is so large that a super-particle tech-
between two unresolved scales need to be modeled properly nique is utilized, in which one computational particle (or parcel)
(Lakehal and Liovic, 2011; Lakehal et al., 2012; Chesnel et al., is used to represent many physical particles. In contrast, for a fully
2011a,b; Vincent et al., 2010; Larocque et al., 2010; Toutant resolved droplet (RD), typically about ten grid points per diameter
et al., 2009a,b; Labourasse et al., 2007; Sirignano, 2005). If a spatial are required, as in Fig. 1(d). The ratio dp =Dx has a significant jump
filter of a constant filtering length is applied to the whole flow field, of about two orders of magnitude between these two limits. Even if
then both the flow properties and the phase-characteristic function AMR is considered, it is impossible to adapt the mesh to such a spa-
need to be filtered (Lakehal and Liovic, 2011; Sirignano, 2005). As a tial jump in a single time step when a resolved droplet is converted
result, an ideal subgrid model would need to capture a variety of to a Lagrangian particle or vice versa. In yet another numerical
multiphase dynamics, such as the formation of sheets, ligaments, setup, if the grid were stretched and the droplet moved from a high
and droplets, droplet dynamics, droplet impact on liquid interface, density grid to a coarse grid region, a transition region of poor res-
and the effect of both resolved and unresolved turbulence on these olution would again be encountered. Thus in all cases the mesh has
pheomena. Although progress has been made in the past decade to be adapted gradually, covering regimes where the grid resolu-
(Lakehal et al., 2012), such comprehensive subgrid models are tion of the droplet lies in between the two extreme limits, as in
exceedingly complex and the current knowledge of many of the Fig. 1(b) and (c). For these two intermediate conditions, dp is only
above phenomena still requires significant improvement for the a few times larger than Dx, and with such a resolution the flow
development of accurate models. inside and outside the droplet is not captured accurately by DNS.
Nevertheless, as the droplets formed in atomization are usually It is therefore reasonable to convert these poorly-resolved droplets
very small, it is easier to only apply subgrid modeling for their to LPP, but because of the larger-than-grid-size droplets the
motion, and the Lagrangian point-particle (LPP) approach is a per- required modeling is much more challenging than in conventional
fect candidate. The LPP model has been widely utilized in case.
particle-laden flows and sprays (Balachandar and Eaton, 2010; The challenge mainly comes from the coupling between the LPP
Apte et al., 2003). In LPP, the particles (or droplets) are viewed as droplets and the resolved fluid flow. The fluid-dynamic force acting
point masses and the flow around individual particles is not on each individual droplet is crucial to LPP models, since it deter-
resolved. To accurately track the particles and the backward effect mines both the droplet motion and the backward effect to the fluid.
to the carrier flow, physical models are needed to calculate the Typically, the calculation of this force requires the undisturbed
instantaneous force acting on the particles. Eventually, in a fluid velocity at the droplet location. For the simple case of a single
so-called Eulerian–Lagrangian simulation, the macro-scale fluid droplet, the ‘‘undisturbed fluid velocity’’ can be considered as the
flow is resolved in an Eulerian framework and the point-particles fluid velocity without the local perturbation induced by the
124 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
Fig. 1. Droplet representation at different grid resolutions: (a) dp Dx=10, (b) dp Dx, (c) dp 4 Dx, and (d) dp 10 Dx.
droplet, or the fluid velocity far away from the droplet location if a
uniform flow is imposed. For the case with many droplets, the u χ=0 u χ=0
undisturbed fluid velocity ‘‘seen’’ by an LPP droplet can be consid-
ered as the fluid velocity when the droplet of interest is absent but
all the other droplets are present (Balachandar, 2009). Because the
χ=1 up χ=0
droplets and the fluid are two-way coupled, not only the fluid up
exerts a fluid-dynamic force on the droplets, but this force is also
exerted back to the fluid with an opposite sign.
Conventionally, the droplet force is applied to the fluid as a
point source located at the center of mass of the droplet (Climent Resolved Drop Lag Point-Particle
and Magnaudet, 2006). Strictly speaking, after the droplet force is
Fig. 2. Schematic representation of a resolved droplet and a Lagrangian point-
added, the undisturbed fluid velocity is not available anymore, particle droplet.
since the fluid flow has been ‘‘perturbed’’. However, for a droplet
that is much smaller than the computational cell the perturbation
induced by each individual droplet to the fluid within a cell is Even for droplets with small Reynolds number, FCM typically
indeed small. As a result, the fluid velocity stored in the cell is a requires four to six grid points per particle diameter (Lomholt
good approximation of the undisturbed fluid velocity. On the other and Maxey, 2003). It is thus inapplicable in atomization simula-
hand, for droplets that are larger than the cell size, if the force is tions, since we need to deal with droplet resolution dp =Dx going
also applied as a point source then the fluid velocity of the cell, continuously from the very small to four grid points.
where the center of mass of the droplet is located, will be influ- The goal of the present study is to propose a multiphase flow
enced significantly and it cannot be considered equal to the undis- model for atomization simulations, where the large-scale inter-
turbed fluid velocity. Furthermore, if the undisturbed fluid velocity faces are resolved by the VOF method and the small droplets by
is not computed correctly, an error will be present in the droplet the LPP model. To focus on the development and validation of
force and it will propagate back to the resolved fluid flow, thus the LPP model, in particular for droplets that are larger than the
leading to serious computational errors (Balachandar, 2009; grid size, we perform the study on a fixed grid. The application
Eaton, 2009). of the LPP model in an octree AMR framework will be considered
It should be reminded that, even though the local perturbation in future works. When an adaptive grid is used, the cell size can
of the flow at the scale of the droplet size is ignored in the LPP be increased (gradually) by at least an order of magnitude, after
model, the LPP droplets and the resolved fluid flow are still the small resolved droplets are converted to LPP droplets. In such
two-way coupled. Since the force exerted on each individual LPP a case, the total computational cost can be reduced significantly.
droplet is subtracted from the resolved fluid flow, the presence For a fixed grid, even though the reduction of computational time
of LPP droplets will influence the resolved flow. However, this is not striking, there are several other important advantages to rep-
influence is only at a spatial scale that is much larger than the dro- resent small poorly resolved droplets by the LPP model. First, the
plet size (Balachandar, 2009). LPP model typically yields more accurate results on droplet
In the conversion from a resolved droplet to an LPP droplet dynamics than keeping the droplets resolved by VOF with poor res-
(RD-to-LPP) the phase-characteristic function v in the region occu- olution (dp < 4Dx). Second, the poorly resolved droplets are usually
pied by the droplet is changed from the value v ¼ 1 of the liquid removed in VOF simulations (Shinjo and Umemura, 2010). An
phase to v ¼ 0 of the gas, as in Fig. 2. Moreover, the local perturba- unaccounted-for removal of the small droplets formed makes the
tion induced by the droplet in the fluid velocity field should be measurements of probability density function (PDF) of droplets
removed carefully, as depicted in Fig. 2, so that the ‘‘undisturbed impossible and also leads to a loss of physics. With the LPP model,
fluid velocity’’ is available to compute the fluid-dynamic force act- these droplets can be kept as LPP droplets. Third, the
ing on the LPP droplet. The inverse conversion (LPP-to-RD) requires poorly-resolved droplets sometimes cause numerical instability
a similar treatment. (which is the motivation to remove them), converting them to
In recent years, more sophisticated LPP models have been pro- LPP improves the numerical stability of the simulation and speeds
posed for larger-than-grid-size solid particles, such as the up the convergence of Poisson solver slightly.
Force-Coupling method (FCM) by Maxey and his colleagues The multiphase flow models, corresponding equations and
(Maxey et al., 1997; Maxey and Patel, 2001; Lomholt et al., 2002; numerical methods are presented in Section ‘Modeling and numer-
Lomholt and Maxey, 2003). Instead of computing the actual ical methods’. A series of multiphase flow problems are considered
fluid-dynamic force exerted on the particle, a Gaussian distribution to validate the approach, and the results obtained by a number of
of external force monopoles and dipoles is applied to mimic the simulations are compared against DNS results and experimental
local perturbation. In such a case, the undisturbed fluid velocity data in Section ‘Results’. In particular, the model is applied to sim-
is not required to model the interphase coupling. However, FCM ulate the gas-assisted atomization experiment by Descamps et al.
is developed based on the multipole expansion of Stokes flows. (2008). This experiment is of particular interest because the drop
Extension to droplets with finite Reynolds number is difficult. trajectories and drop flying angle statistics are measured. The
Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 125
multigrid solver, probably because of the small time step dup u ~ g up qg Du~ g C m qg Du~ g dup
¼ /þ þ þ g; ð6Þ
constraint. dt sp qp Dt qp Dt dt
neglected in the present LPP model. For the gas assisted atomiza- ~ g , we propose to calcu-
In order to obtain a better estimate of u
tion configuration of interest here, the small droplets formed in late f p as
atomization are dispersed downstream of the liquid jet. As a result,
Np
X
their total volume fraction is quite low and the inter-droplet inter-
fp ¼ F fp;i Gðx xp;i Þ; ð14Þ
action is of secondary importance. Nevertheless, there exist
i¼1
atomization applications where the effects of inter-droplet interac-
tion and secondary breakup are important (Apte et al., 2003), it is where N p is now the total number of LPP droplets in the flow field.
interesting to extend the present model to include these effects in The Gaussian function Gðx xp;i Þ is a numerical representation of
future works. the LPP droplet coupling force (Maxey et al., 1997)
3=2
GðxÞ ¼ ð2pL2 Þ expðjxj2 =2L2 Þ; ð15Þ
Momentum coupling between LPP and resolved flow
where L controls the size of the region where the force should be
The force exerted by the resolved flow on the LPP droplets is distributed. Eq. (14) is similar to that given by Maxey et al.
referred to as forward coupling and its computation is based on (1997), but the interpretation is different. In the present model, L
the resolved flow properties, Eq. (6). Due to Newton’s third law, is viewed as the length scale of a spatial filter of the coupling force
the force exerted on the LPP droplets needs to be subtracted from at the LPP droplet location. The parameter L is chosen to be larger
the resolved flow and is referred to as backward coupling. This than the actual droplet size. Depending on the volume fraction of
forcing term is represented by f p in the resolved flow momentum the LPP droplets, L is picked between 5 and 10 dp . Therefore, the
equation, Eq. (1). th
coupling force F fp;i corresponding to the i LPP droplet is smoothly
In the conventional LPP model, the computational cell is usually
distributed within a region larger than the droplet size, rather than
much larger than the LPP droplet size and many droplets are con-
concentrated at the droplet location. At every LPP droplet location,
tained in each cell. The forcing term f p is then computed by a vol- the contribution of the force from the present droplet to the local
ume average over the LPP droplets resolved flow is small, so that ug computed from Eq. (1) can be used
N as an approximation of u ~ g in Eq. (6).
1 X p;c
fp ¼ F fp;i ; ð12Þ There are several advantages for treating the coupling force in
V c i¼1 the above way. First, the undisturbed fluid velocity is directly given
by the resolved flow momentum equation to calculate the droplet
where V c and N p;c are the cell volume and the number of droplets in
force. Second, the integration of Eq. (14) over the whole domain
that cell. Note that only the force due to fluid-LPP-droplet interac-
yields the total force exerted on all the LPP droplets and the total
tion needs to be subtracted from the momentum equation
momentum of the combined RF-LPP system is conserved. Third,
dup;i since L is determined by the droplet diameter and not by the cell
F fp;i ¼ mp;i g : ð13Þ size, the coupling between RF and LPP droplets does not depend
dt
on the grid. This feature is quite important if the LPP model is cou-
Although there exist issues of coupling LPP back to the resolved pled with AMR, because the calculation of the force on the LPP dro-
flow due to the number density fluctuation (Sundaram and Collins, plets and the backward effect on the resolved flow will not be
1996; Ling et al., 2010), the handling of two-way coupling is rela- affected when the mesh adapts itself from the well-resolved dro-
tively easy. Since the fluid velocity corresponds to an average value plet regime, Dx < 0:1dp , to the conventional LPP model, Dx > 10dp .
over the cell, and the cell contains a large number of LPP droplets, The present treatment of the coupling force will only capture
then their influence is automatically filtered out (but the effects of the backward effect of the LPP droplets on the resolved flow on
the ambient flow and other LPP droplets outside the cell are length scales larger than L. The dynamics due to fluid-droplet
retained). As a result, the stored fluid velocity can be used directly interaction on scales smaller than L, such as the shear layer near
as the undisturbed fluid velocity in Eq. (6). the droplet interface, the wake of the droplet and the interaction
However, in the present LPP model the droplets are bigger than with the resolved flow, will be missed.
the cell size, the two-way coupling becomes more complicated and Finally, since the LPP droplets do not in general coincide with
in particular Eq. (12) is not valid any more. What’s even worse is the locations at which the fluid velocity is stored, interpolation is
the fact that if the force F fp;i is applied to the cell where the mass required to compute ug at the LPP droplet location, and a
center of the LPP droplet is located, or is distributed to all the cells tri-linear interpolation scheme is used in the present study (Apte
occupied by the LPP droplet, the resolved flow will be affected et al., 2003; Ling et al., 2010). When the spatial variation of the
significantly. resolved flows is significant over the LPP droplet size, then picking
For this reason, we retain Eq. (6) and, as suggested by Maxey the fluid velocity at the LPP droplet center is not sufficient and the
et al. (1997), write the resolved velocity ug as the sum of two con- Faxén correction is often used to generalize the force models
tributions, ug ¼ u ~g þ u ^ g . The first term u
~ g represents the undis- (Faxén, 1922). Furthermore, if the LPP droplet moves in a shear
turbed fluid velocity, the second one u ^ g is the velocity flow, then the shear lift force will be induced (Legendre and
disturbance induced by the droplet under examination, a term Magnaudet, 1997). Nevertheless, the LPP model is only applied to
which is filtered out in the conventional LPP model. small droplets in the present study, the effect of inhomogeneous
When there are only a few LPP droplets in the flow field, their ambient flows on the droplet force is expected to be generally
contribution to the momentum equation of the resolved flow is small and thus the Faxén and lift forces are neglected.
small, and a simple solution is to neglect the coupling force term
f p in Eq. (1). This is the one-way coupling approximation for which Two-way conversion between LPP and resolved droplet
^ g ¼ 0 and u
u ~ g ¼ ug . It has been found in previous works that the
effect of the coupling force on changing the fluid momentum is The flow chart of the combined algorithm in Fig. 3, shows the
related to the mass fraction of dispersed phase (Crowe et al., tight coupling between the two solvers for the resolved flow and
1998; Ling et al., 2013), hence in the one-way coupling assumption the LPP droplets. In atomization simulations that consider the jet
the error in u ~ g increases with the mass fraction of the LPP droplets formation as well, the initial flow is completely resolved with the
(Eaton, 2009). VOF method, since small droplets are not yet formed. Once the
128 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
liquid jet starts to atomize, the droplets away from the interface for icell =1,Ncell do
are converted to LPP droplets. Some of the LPP droplets may follow Identify separated droplets and generate RD array
Tagging
trajectories that bring them back toward the interface. These dro- Perform droplet calculation (droplet volume, etc)
end for
plets may be converted back to resolved droplets. The conversion
for iRD =1,NRD do
algorithm is summarized in Fig. 4.
Check conversion criteria
if Droplet iRD qualified for LPP
Conversion criteria RD-to-LPP
Change non-zero color function to zero
The criteria to determine whether a droplet should be repre- Replace perturbed flow field by undisturbed flow field
Conversion
Add droplet iRD into LPP array and store droplet position and
sented by the LPP model or remain resolved depend on its shape velocity
and position. end if
The present LPP force model assumes the droplet to be spheri- end for
cal. When the droplet Weber and Reynolds are large, or when the for iLPP =1,NLPP do
droplet Capillary number Cap ¼ Wep =Rep becomes large the dro- Check conversion criteria
if Droplet iLPP not qualified for LPP
plet may deform significantly and the LPP model will introduce LPP-to-RD Reconstruct color function for the droplet
an error in its motion. However, if the droplet diameter is less than Conversion Change velocity within the droplet
four to six grid spacings, then the shape and the dynamics of the Remove droplet iLPP from LPP array
end if
droplet are not accurately represented even by the VOF method.
end for
Therefore, droplets with moderate Weber and Reynolds numbers
should also be converted to LPP droplets. In previous atomization Fig. 4. Algorithm for conversation between resolved droplets (RD) and Lagrangian
simulations these poorly resolved droplets, with dp 6 4 Dx, were point-particle (LPP) droplets.
removed to improve the stability of the solver (Shinjo and
Umemura, 2010). Here, instead of simply removing them, we con-
vert them to LPP droplets, with a positive effect on the numerical =0
=0 Reconstructed
stability and on the mass conservation, and with better droplet u undisturbed u
statistics during the atomization process. flow field
For convenience of the calculation, the droplet volume V p is u p=
used instead of its diameter dp to define the droplet size. The aspect (∫u dV)/Vp
ratio c is also estimated in order to avoid converting to LPP dro- =1 RD-to-LPP
plets thin ligaments and sheets. Therefore, only when V p is smaller
than a critical value, say V crit ’ ð4DxÞ3 , and c is close to one, then
the droplet is qualified for the LPP model. Resolved Drop (a) Lag Point-Particle
Furthermore, a location criterion is used to check if the droplet
is away from the liquid jet interface. Since the current LPP model
does not include either the formation of a droplet or a droplet =0 =0
u u
impact on the interface, only droplets that are at a certain distance
away from the liquid jet interface can be converted. The distance is
typically chosen to be the droplet diameter dp . up+u’p =0
The overall conversion criteria can be expressed as =1 LPP-to-RD up
fLPP Qualifiedg ¼ fV p < V cut g && fjcp 1j < c g && fxp 2 Rai g;
ð16Þ
Resolved Drop (b) Lag Point-Particle
where c is the tolerance for the aspect ratio, and Rai is the region
away from the interface. Fig. 5. Conversion between a resolved droplet (RD) and a Lagrangian point-particle
(LPP) droplet. The fluid flow field for the LPP droplet is slightly perturbed due to the
coupling force.
RD-to-LPP conversion
For the conversion from resolved droplets to LPP droplets, the
separated liquid structures must be first identified by tagging the
color function field. The tagging approach of Herrmann (2010) is two times the droplet diameter (blue1 area of Fig. 5(a)). For exam-
employed here, and cells that are attached to each other, with ple, for the velocity component u along the x-direction, a linear
respect to the liquid phase, will have the same tag number. interpolation is performed between the two values uði1 1; j; kÞ
During the tagging process, calculations of the droplet properties, and uði2 ; j; kÞ, where i1 and i2 are the indexes of the left and right
including its volume, aspect ratio, location and velocity of the cen- boundary cells of the cubic region. The reconstructed velocity field
ter of mass, are also performed. is globally divergence-free if the velocity on the surface of the
A schematic representation of the conversion from RD to LPP is interpolation region is divergence-free. The field is also
shown in Fig. 5(a). In the conversion, the droplet volume V p and divergence-free locally if the divergence components
velocity up are first computed by an integration over the region @u=@x; @ v =@y, and @w=@z are constant within the region. Unless
X occupied by the droplet (gray area of Fig. 5(a)), namely the particle Reynolds number is very small, a cubic region with
R R
V p ¼ X v dV and V p up ¼ X u v dV, the phase-characteristic func- an edge size twice the droplet diameter is sufficient to reconstruct
tion v is then changed from the value 1 to 0. the undisturbed flow field in a satisfactory way. The reconstruction
Moreover, the perturbed flow field in the same region should be of the undisturbed flow field is necessary even for the simple
replaced by the undisturbed flow, so that the LPP droplet sees the approach of removing the small poorly-resolved droplets from
proper undisturbed fluid velocity. This is done by interpolating
each component of the fluid velocity along the corresponding coor-
dinate from the surface into the interior of a cubic region centered 1
For interpretation of color in Figs. 5, 10, 11, 14 and 16, the reader is referred to the
at the particle location. The edge length of the cubic region is about web version of this article.
Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 129
the simulation, since which minimizes the influence of the removal The second term on the right hand side is the momentum of the
of droplets to the resolved flows. It has been observed in the previ- perturbed flow field, which is approximated by assuming a virtual
ous work that if the small liquid structures are replaced by gas volume of fluid, aV p , moving with the droplet velocity,
with the same velocity, the interface instability in the resolved
flow is modified (Shinjo and Umemura, 2010). Z
Since in the RD-to-LPP conversion, the region originally occu- qg ðug u~ g ÞdV aqg V p ðup u~ g Þ: ð18Þ
Vp
pied by the liquid will be filled with the gas phase, it might seem
that additional gas is added to the system and mass conservation Then the velocity to be used in the cells of the resolved droplet
is violated. As a matter of fact, the added gas is actually considered in the LPP-to-RD conversion can be expressed as
as a ‘‘ghost fluid’’ and is used only for the droplet force calculation.
When an LPP droplet is converted back to a resolved droplet, the !
qg
ghost fluid will be taken out from the system. Therefore, the ghost u0p ug ¼ 1 þ a ðup ug Þ: ð19Þ
fluid should always be excluded from the calculation of the total qp
mass of the system.
The physical meaning of a is the ratio between the volumes of
Furthermore, the LPP droplet represents not only the momen-
the virtual fluid and the droplet. However, it is not easy to get a
tum of the droplet, but also the momentum of the unresolved per-
precise value of a. It is known that, in order to accelerate a particle,
turbed flow around the droplet. The latter can be viewed as the
additional forces, i.e., the added-mass and history
momentum of a virtual volume of fluid that moves with the LPP
(viscous-unsteady) forces, are needed to accelerate the ambient
droplet. Therefore, although the perturbed flow around the droplet
fluid around the particle. For the added-mass effect, it is considered
is removed in the RD-to-LPP conversion, the corresponding
that the ratio between the volumes of the virtual fluid to be accel-
momentum is kept. As a result, the momentum of the overall sys-
erated through the inviscid mechanism and the particle is the
tem is conserved.
added-mass coefficient C M , which is equal to 0.5 for a sphere in
incompressible flows. Due to finite viscous diffusion time scale,
LPP-to-RD conversion
the history force usually needs to be expressed in integral form
Several conversion criteria can be devised to examine if an LPP
(Basset, 1888; Mei and Adrian, 1992). Nevertheless, it has been
droplet is still qualified for the LPP model, for example the droplet
shown by Ling et al. (2013) that, if the particle and ambient fluid
volume can increase due to collision and merging of droplets or to
acceleration time scales are much larger than the viscous diffusion
phase change. These physical processes are not considered in the
time scale, the history force can also be expressed as non-integral
present LPP model, where the only condition for converting an
form like the added-mass force, and a viscous-unsteady coefficient
LPP droplet to a resolved one is that the droplet position is very
C v u similar to the added-mass coefficient C M can be derived as
close to the liquid jet interface.
A schematic picture of the conversion process is shown in
Fig. 5(b), and the algorithm is described in Fig. 4. First, a spherical 0:75 þ 0:105Rep
C v u 8:51 : ð20Þ
droplet is rebuilt around xp , by updating properly the color func- Rep
tion in the cells that will be occupied by the resolved droplet. The viscous-unsteady coefficient C v u , can be considered as the
The ghost fluid that was added in the inverse RD-to-LPP conversion ratio between the volumes of the virtual fluid to be accelerated
is now taken out. through the viscous mechanism and the particle. It is shown that
In the same cells containing the new resolved droplet the veloc- C v u decreases with the droplet Reynolds number Rep . When Rep
ity field needs to be updated to take care of the LPP droplet veloc- varies from 10 to 100, C v u decreases from 1.53 to 0.96. The excess
ity. In principle, the local perturbed flow inside and outside the momentum added through u0p is to mimic the effects of the
droplet should be reconstructed, this being the inverse process of
added-mass and history forces on accelerating the surrounding
the removal of the perturbed flow in the RD-to-LPP conversion.
fluid around the droplet. Therefore, a can be estimated as the
The reconstruction of the local perturbed flow is quite a difficult
sum of C M and C v u and thus depends on the droplet Reynolds num-
task, except in the limit of zero Reynolds number for which an ana-
ber as well. In an actual computation, the resolution of RD is usu-
lytical solution exists (however, in that case the perturbed region is
ally 4–6 grid cells across the diameter, and the local perturbed flow
infinitely large). Nevertheless, if one naively ignores the effect of
may not be computed very accurately. As a result, a usually needs
the local perturbed flow and only changes the velocity inside the
to be adjusted slightly. For simplicity, a constant of a ¼ 4 is used in
droplet to be the LPP droplet velocity up , then it can be easily
the present study, which has ben shown to yield a smooth droplet
shown that the droplet velocity will have a significant transition
velocity transition in the LPP-to-RD conversion.
jump right after the LPP-to-RD conversion. This is simply due to
the fact that a portion of the momentum of the droplet given in
the conversion is transferred to the surrounding flow and used
for the local perturbed flow development. Results
As described in the RD-to-LPP conversion, the LPP droplet
actually contains the momentum of both the droplet and the vir- To validate the combined multiphase flow model where the
tual fluid moving with the droplet. In the naive conversion smallest droplets are represented by the LPP model, four test prob-
approach described above, only the portion of droplet is included. lems with increasing complexity are conducted: settling droplet,
In order to compensate the perturbed flow that is not recon- droplet in a lid-driven cavity, pulsed jet atomization, and
structed, the momentum of the virtual fluid also needs to be gas-assisted atomization. The combined model considers the
added back. Therefore, the velocity in the cells occupied by the point-particle approximation for small droplets, that introduces
resolved droplet should be changed to u0p for the purpose of com- errors even when the droplets are spherical. Nevertheless, these
pensating the momentum in the perturbed flow field around the errors should be small and should not influence the features of
droplet, i.e., interest to be captured by atomization simulations, such as droplet
Z size PDF and droplet flying angle PDF. To validate the combined
qp V p ðu0p u~ g Þ ¼ qp V p ðup u~ g Þ þ qf ðug u~ g ÞdV; ð17Þ model, its numerical results are compared with experimental data
Vp or DNS, depending on the test problem.
130 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
The combined model is first tested with the simulation of a liq- ql (kg/m3) qg (kg/m3) ll (Pa s) lg (Pa s) dp (m) r (N/m) g (m/s2)
uid droplet settling in a large rectangular cuboid filled with a sta-
1000 10 10 3
104
10 4 0 9.8
tionary gas, see Fig. 6(a). The relevant physical properties of the
two phases and other parameters are given in Table 1. The dimen-
sions of the tank are Lx ¼ Lz ¼ Ly =2 ¼ 103 m. Gravity is along the
dp =h ¼ 8; 16, when the droplet is sufficiently resolved. When the
negative y-direction. The center of the droplet is initially at
mesh resolution is low, dp =h ¼ 2; 4, DNS yields errors of 16.2%
(0:5 Lx ; 0:5 Ly ; 0:5 Lx ) and reaches the settling velocity at about time
and 7.0% in the terminal velocity, respectively. Therefore, in this
t ¼ 0:03 s. The droplet terminal velocity is about v t ¼ 0:04 m/s
case LPP model yields more accurate results than DNS of a poorly
with a corresponding droplet Reynolds number Ret 0:4. The dro-
resolved droplet.
plet remains about spherical during the settling due to viscous
Though there is no experimental data for this specific test case,
effects, and surface tension is therefore not included in this
there are empirical correlation for the terminal settling velocity to
simulation.
compare with. For a droplet settling in an unbounded domain, the
Note that the standard drag correlation Eq. (10) is for LPP in an
terminal velocity obtained through the standard drag (Clift and
unbounded domain. Here, the tank width and depth are only 10
Gauvin, 1970) is about v t ¼ 0:0493 m/s. This value is considerably
times the droplet diameter and, as a result, the tank wall has a
larger than the simulation results, implying that the wall effect is
noticeable retarding effect on the settling velocity v t . The relation
significant and should be considered. As a function of the
between the terminal settling velocity v t and the
Reynolds number, in the two limits Rep ! 0 and Rep ! 1, the ter-
diameter-to-width ratio dp =Lx has been studied by many research-
minal velocity, including the wall retarding effect, can be computed
ers, typically in a cylindrical tube with a circular cross section (Di
as v t;0 ¼ 0:0393 m/s and v t;1 ¼ 0:0478 m/s (Di Felice, 1996). Both
Felice, 1996). To include the wall effect, the correlation of Di
DNS and LPP results, with the exception of the lowest resolution
Felice (1996)
dp =h ¼ 2, are bounded by these two limits. In particular, since the
2:7
vt 1 dp =Lx Reynolds number based on the droplet terminal velocity is about
g¼ ¼ ; ð21Þ Rep ¼ 0:4, we expect and find that our results are closer to the
v t;unbounded 1 0:33 dp =Lx
lower limit. With the correlation of Di Felice (1996) the terminal
is used to correct the drag calculation, where g is defined by the velocity is about v t ¼ 0:0406 m/s, which again agrees reasonably
ratio between the terminal settling velocity in bounded and well with our simulations. The small discrepancy may be also
unbounded domains. This correlation is here used to correct the due to the different shape of the domain cross section in the pre-
quasi-steady force in the equation of motion (6) in the following sent study and in Di Felice (1996).
way
dup u ~ g up / qg Du
~ g C m qg Du~ g dup Lid-driven cavity
¼ þ þ þ g: ð22Þ
dt sp g qp Dt qp Dt dt
In the second test problem, we investigate the droplet motion in
In DNS without the LPP model and with cubic cells of side h, a 3D lid-driven cavity. The cavity is a cube of size L and the flow is
four different resolutions, dp =h ¼ 2; 4; 8; 16, are used to resolve driven by a moving lid on the upper boundary with velocity U 0 ; the
the droplet. With the LPP model there is only one droplet in the Reynolds number of the cavity is Rec ¼ 100. The boundary condi-
tank, hence the one-way coupling approximation is made and f p tions along the x- and y-directions are no-slip, periodic boundary
~ g and Du
is ignored in Eq. (1). As a result, u ~ g =Dt are equal to zero. conditions are considered along the third z-direction. The relevant
The LPP model result is simply obtained by integrating Eq. (22) physical properties of the two phases and other parameters are
and thus it is independent of the cell size. The results for the given in Table 2. Simulations are first run without any droplet, then
y-velocity of the droplet with only the VOF method or the LPP the induced flow remains 2D, moving towards the steady state
model are shown in Fig. 6(b). It can be observed that the LPP result condition shown in Fig. 7. In Fig. 7(b), the u component of the
agrees well with the DNS results obtained with the finest meshes, velocity (x-direction) along the vertical centerline y ¼ z ¼ L=2,
0
LPP
DNS,dp/h= 2
DNS,dp/h= 4
-0.01 DNS,dp/h= 8
DNS,dp/h=16
-0.02
v (m/s)
-0.03
-0.04
-0.05
0 0.005 0.01 0.015 0.02 0.025 0.03
t (s)
(a) (b)
Fig. 6. A liquid droplet settling in a stationary gas tank: (a) problem setting, (b) settling velocity as a function of time.
Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 131
N=128
N= 64
Ghia
(a)
(b)
Fig. 7. Single phase results of the lid-driven cavity problem: (a) velocity vector plot at z ¼ L=2, (b) u at y ¼ z ¼ L=2 and v at x ¼ z ¼ L=2.
Table 3
The effect of the paremeter a on the droplet velocity in LPP-to-RD conversion. The superscript 1 and 2 denote the droplet velocity right before and 6 steps after the LPP-to-RD
conversion. The droplet velocities for LPP-to-RD conversion with different a are represented by u2p;RD ðaÞ; while the LPP droplet at the corresponding time without a conversion is
denoted by u2p;LPP . The lid velocity U 0 ¼ 3:125 m/s and the undisturbed fluid velocity at the LPP droplet location before conversion is (0:1839U 0 ; 0:02928U 0 ; 0).
-0.2 0.7
v (m/s)
-0.4 RD-to-LPP 0.3 RD-to-LPP
-0.45 0.2
-0.5 0.1
-0.55 0
LPP-to-RD LPP-to-RD
-0.6 -0.1
0.001 0.00105 0.0011 0.00115 0.0012 0.001 0.00105 0.0011 0.00115 0.0012
t (s) t (s)
(a) (b)
Fig. 8. Time evolution of the velocity components when the droplet initial velocity is the same as the local fluid velocity: (a) u, (b) v.
0.1 0.3
DNS (RD)
0 LPP DNS (RD)
RD-LPP-1 0.2 LPP
-0.1 RD-LPP-2 RD-LPP-1
RD-LPP-2
-0.2 0.1
v (m/s)
u (m/s)
LPP-to-RD
-0.3
-0.4 0
LPP-to-RD
-0.5
-0.1
-0.6
-0.7 -0.2
0.001 0.0011 0.0012 0.0013 0.001 0.0011 0.0012 0.0013
t (s) t (s)
(a) (b)
Fig. 9. Time evolution of the velocity components when the droplet initial velocity is different from the local fluid velocity: (a) u, (b) v.
indeed quite small. Furthermore, this error magnitude is also con- critical to handle properly the local flow field in the conversion
sistent with the droplet low resolution, dp =h ¼ 4. process.
When we consider the conversion model with the initial veloc- We also consider the case with multiple droplets seeded in the
ity up1 we observe that the droplet at about time t ¼ 1 ms is on the cavity. In total 183 monodisperse droplets are seeded randomly
lower boundary of the slab with a negative y-component of the after a steady state flow is reached in the cavity with a single
velocity. The LPP particle is quickly converted to RD, then after a phase. The mass fraction of the seeded droplets, defined by the
transient it changes the sign of this velocity component and finally ratio of their mass to the total mass of liquid droplets and gas, is
it enters the slab region where it is converted back to LPP. A similar about 0.75. In this case the mass fraction is large, and two-way
discussion holds for the initial velocity up2 . With the conversion coupling simulations are performed with the force coupling model
model described in Section ‘Two-way conversion between LPP described in Section ‘Momentum coupling between LPP and
and resolved droplet’, the droplet velocity has a smooth transition resolved flow’. The size of the parameter that controls the force dis-
after the conversions for both initial velocities up1 and up12 . In tribution is L ¼ 10 dp .
Figs. 8 and 9 this model is represented by the RD-LPP-1 lines, with Fig. 10 shows the droplet distribution inside the cavity at times
the conversion instants marked by an arrow. If the perturbed flow t ¼ 1:04; 1:08; 1:12 ms. When the droplets are resolved by the VOF
field is not reconstructed properly in the LPP-to-RD conversion, method the interfaces are plotted as the intermediate contour level
then remarkable transition jumps are observed in the droplet C ¼ 0:5, and are indicated by a red color. The droplet distribution
velocity after the conversion (RD-LPP-2 lines). The transition jump represented by the LPP model is given by the green spheres. The
is more significant when the droplet is not in velocity equilibrium DNS and LPP results overall match well on the droplet location.
with the local fluid. The differences, shown in the images on the third column, mainly
As the drag force tends to bring the droplet in velocity equilib- occur when the droplets are subject to collisions or serious defor-
rium with the fluid, the droplet velocity may slowly change toward mations, as near the top of the cavity where the shear is strong.
the correct value, as in Fig. 8, but the transient process can be quite This indicates the need for a future improvement of the model to
long. However, if the initial error in the conversion causes a signif- take into account both droplet–droplet and droplet–wall interac-
icant deviation of the droplet trajectory, the droplet velocity will tions, such as incorporating the collision model of Schmidt and
never go back to the correct value, as in Fig. 9(b). Therefore, it is Rutland (2000).
Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 133
(a)
(b)
(c)
Fig. 10. Droplet distributions in the cavity at times: (a) t ¼ 1:04 10 , (b) t ¼ 1:08 103 , and (c) t ¼ 1:12 103 s.
3
It is also important to point out that the results obtained with Table 4
the one-way and the two-way coupling models are indeed very Parameters for the pulsed-jet atomization problem.
close. Therefore, the one-way coupling with the present conversion
ql =qg ll =lg Rejet Wejet
model seems to be adequate even when the LPP droplet mass frac-
tion is quite large. Nevertheless, further validation of the coupling 20 20 800 4000
droplets in the flow is small during all the simulation, and based
Pulsed-jet atomization
on the knowledge of the previous test, the one-way coupling model
is used. Since the amount of liquid injected into the domain
Next we consider the atomization of a dense cylindrical liquid
increases with time, the total number of droplets and the number
jet which is injected into a stagnant gas tank, to examine the capa-
of those represented by LPP model also increase. The maximum
bility of the present model to simulate multiphase flows with very
number of LPP droplets is about 6000.
complex interfaces. The primary breakup of a liquid jet injected
The simulation of the pulsed-jet problem is conducted on the
into stagnant gas has been investigated by many previous works
CALMIP machine, with a typical CPU time on 160 processors of
in detail, see for example (Ménard et al., 2007; Lebas et al., 2009;
245 min for the combined model and 253 min for DNS.
Shinjo and Umemura, 2010). Discussions of the primary breakup
Snapshots of the interface configuration in the two simulations
mechanisms of the liquid jet can be found in the above reference.
at time t ¼ 9 Djet =U jet are shown in Fig. 11(a) and (b), with the VOF
In the present paper, we mainly focus on testing the present model.
interface represented in green and white for the contour level
The jet diameter is denoted by Djet and the tank dimensions are
C ¼ 0:1, in order to make small droplets clearly visible. The orange
Ly ¼ Lz ¼ 4 Djet and Lx ¼ 10 Djet . The injection is along the positive
spheres represent the liquid droplets of the LPP model, and their
x-direction. The inflow velocity u is modulated sinusoidally, radius length is such to give the actual droplet volume. The two fig-
u ¼ U jet ð1 þ 0:05 sinð20 p tÞÞ, to promote the growth of primary ures are in good agreement with each other, thus confirming the
shear instabilities. The physical properties and parameters in capability of the combined model to capture atomization.
non-dimensional form are listed in Table 4, where The primary breakup of the pulsed liquid jet can be clearly seen
Rejet ¼ ql U jet Djet =ll and Wejet ¼ ql U 2jet Djet =r. in Fig. 11. Liquid sheets are formed at the front of the liquid jet and
On the left and right boundaries, inflow and outflow boundary also on the liquid core surface. At the front, a umbrella-shape sheet
conditions are considered, respectively, while free-slip conditions is formed due to the impingement of the liquid jet against the stag-
are applied on the other four boundaries. The grid is composed nant gas and the resulting lateral liquid spread. On the liquid core
of cubic cells of side h ¼ Djet =64. Two simulations are performed surface, it is seen that the perturbation near the inlet remains
for this problem: the first one is DNS with the VOF method, the sec- almost sinusoidal, but as it is advected downstream, the interface
ond one with the combined method, with the small droplets folds and thin liquid sheets are formed. The sheets incline toward
described by the LPP model. Since the mass fraction of the LPP the upstream direction, looking like an umbrella. As the sheets are
134 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
Fig. 11. Snapshots of the liquid–gas interface of the pulsed-jet atomization at time t ¼ 9 Djet =U jet : (a) DNS, (b) the combined model, where the orange droplets are from the
LPP model.
stretched, they break up into thin ligaments. The ligaments in present study, because the drop trajectories and drop flying angle
sequence break into small droplets due to the Plateau-Rayleigh statistics are measured in the experiment and thus can be used
instability (Shinjo and Umemura, 2010). to validate the capability of the present combined model in accu-
The size distribution (PDF) of the droplets formed in the rately capturing the drop dynamics. Up to the authors’ knowledge,
pulsed-jet atomization at different times is shown in Fig. 12. the experiment by Descamps et al. (2008) is the only one in the lit-
Since more and more liquid is injected into the domain and then erature that provides detailed measurements of drop trajectories.
atomized to small droplets as time evolves, the droplet Sauter
mean decreases with time, while the number of small droplets
Simulation setup
increases. It can be seen that the droplet size distributions obtained
The simulation setup is shown in Fig. 14. The atomizer consists
by DNS and the combined model are in good agreement. In partic-
of two planar jets: on top the faster gas jet with injection velocity
ular, the combined model captures the peak of the PDF, while the
U g , and on bottom the slower liquid jet with injection velocity U l .
difference of the Sauter means with the DNS results is less than
The inlet height for both phases is the same, and is denoted by
10%. When the LPP model is introduced, the number of tiny dro-
H. As a function of this characteristic length, the width, height
plets is significantly higher than the DNS results without the LPP
and thickness of the computational domain are
model, see Fig. 12. The discrepancy may be due to the erroneous
Lx =8 ¼ Ly =4 ¼ Lz ¼ H, respectively. The domain is subdivided with
dynamics of small droplets when they remain to be resolved by
cubic cells of side h ¼ H=64 for the coarse mesh and h ¼ H=128
VOF. Small fragments of VOF-represented liquid tend to be recon-
for the fine one.
structed and advected erroneously, so that these small fragments
At the inlet, a separator plate, indicated by the blue color in
stay where they are formed they are easily swept by and merged
Fig. 14, has been positioned to separate the gas and liquid flows.
into the moving bulk liquid jet. If these small fragments are con-
The need for such a plate to accurately simulate atomization has
verted to LPP, then they can be advected away from the liquid jet
been shown in previous papers (Fuster et al., 2013). The width
and remain conserved. As a result, less small droplets are seen in
and height of the separator plate are lx ¼ H=4 and ly ¼ H=32,
the pure VOF simulation than the combined VOF-LPP simulation.
respectively. Inflow and outflow boundary conditions are applied
The forward RD-to-LPP process and the backward LPP-to-RD
along the x-direction, slip conditions along the vertical y-direction,
conversion, from the ligament breakup to the droplet merging back
periodic boundary conditions along the z-direction. A snapshot of
in the liquid jet, are shown in Fig. 13. In particular, a ligament
the liquid–gas interface resolved by the VOF method (in gray)
formed at the edge of the sheet is breaking due to
Plateau-Rayleigh instability in Fig. 13(a). In Fig. 13(b), it can be and by the LPP model (in orange) is also shown in Fig. 14.
The relevant physical properties of the two phases and other
seen that the initially formed droplet has a rather irregular shape.
Under the action of surface tension, the droplet becomes closer to a parameters are given in Table 5. The properties of the gas and liq-
uid phases are chosen in such a way to guarantee that the impor-
spherical shape in Fig. 13(c). Then, since the droplet satisfies both
tant dimensionless parameters are as close as possible to the
the size and aspect-ratio criteria and is enough away from the jet, it
corresponding value of the experiment by Descamps et al. (2008).
is converted from RD to LPP, as shown in Fig. 13(d). The LPP droplet
In particular, Table 6 shows that the Reynolds and Weber num-
moves for a while in the gas, then driven by the recirculation gen-
erated on the inside of the umbrella-shape front, the droplet is car- bers of the liquid inflow, Rel ¼ ðql U l HÞ=ll and Wel ¼ ðql U 2l HÞ=r,
ried back to the liquid jet. When the LPP droplet is getting close to and the momentum ratio between the gas and liquid inflows,
the interface, it does not satisfy the criteria of LPP and then under- M ¼ ðqg U 2g Þ=ðql U 2l Þ, are very close to the experimental value. On
goes a backward conversion from LPP to RD in Fig. 13(e). Finally it the other hand, while the numerical Reynolds number based on
is reabsorbed by the liquid jet in Fig. 13(f). The present results indi- the gas inflow boundary layer thickness dg , Reg;d ¼ ðqg U g dg Þ=lg ,
cate that the two-way conversion scheme is capable to capture is also very similar, the corresponding Weber number,
complex droplet dynamics in atomization. Weg;d ¼ ðqg U 2g dg Þ=r, is about twice as big. Finally, the numerical
Reynolds number of the gas jet is fairly large,
Gas-assisted atomization Reg ¼ ðqg U g HÞ=lg 8000, but still only about 42% of the experi-
mental value. For such high Reynolds numbers, it is extremely
The combined model is finally applied to simulate the expensive to resolve all turbulent scales in the gas flow, and we
gas-assisted atomization experiment conducted by Descamps do not expect that the current mesh resolution can capture the
et al. (2008). This experiment is particularly interesting to the smallest turbulent eddies. Therefore, the term ‘‘DNS’’ here refers
Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 135
10 3 10 3 10 3
DNS DNS DNS
Model Model Model
10 2 10 2 10 2
counts
counts
counts
10 1 10 1 10 1
10 0 10 0 10 0
0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02
d d d
(a) (b) (c)
Fig. 12. Size distribution and Sauter mean of droplets formed in pulsed-jet atomization at times (time units Djet =U jet ): (a) t ¼ 8; d32;DNS ¼ 2:60 mm; d32;DNS ¼ 2:88 mm, (b)
t ¼ 8:5; d32;DNS ¼ 2:55 mm; d32;DNS ¼ 2:66 mm, (c) t ¼ 9; d32;DNS ¼ 2:33 mm; d32;DNS ¼ 2:49 mm.
(d ) (e ) (f)
Fig. 13. The two-way RD-to-LPP and LPP-to-RD conversions for a droplet initially formed from a ligament breakup and later reabsorbed by the liquid jet, at times (time units
Djet =U jet ): (a) t ¼ 2:0, (b) t ¼ 2:2, (c) t ¼ 2:4, (d) t ¼ 2:6, (e) t ¼ 2:8, (f) t ¼ 3:2.
Table 5
Parameters for the gas-assisted atomization problem.
Lz=H
Lx=8H
Periodic ql qg ll lg r Ul Ug H dg (m)
(kg/m3) (kg/m3) (Pa s) (Pa s) (N/m) (m/s) (m/s) (m)
c=0 Slip wall
1000 10 103 104 0.2 8.3 0.1 0.01 0.0011
Inflow LPP
Separator plate
Gas H interface Outflow
Table 6
Important dimensionless parameters of the present simulation and the experiment by Descamps et al. (2008) of gas-assisted atomization.
Experiment 833 58.8 1000 1.45 69.1 960 5.68 1:69 104
Simulation 100 10 1000 1.45 68.9 913 11.0 8:30 103
General behavior instability, as shown in Fig. 16(f). Finally, the atomized droplets
The liquid jet atomization assisted by the fast gas stream is are advected downstream and leave the computational domain.
shown in Fig. 16 at different times. The gray and orange colors indi-
cate the interface resolved by the VOF method and the small dro-
plets represented by the LPP model, respectively. The vector plot Droplet dynamics and flying angle statistics
on the background shows the flow velocity, where the vector mag- Due to the strong interaction between the wavy interface and
nitude is indicated by both length and color of the arrow. The flow the gas stream, the formation and dynamics of the droplets are
velocity within the liquid is much smaller than in the gas, and the rather complex. The droplets are generated along the streamwise
length of the arrows is too small to be appreciated. The simulation direction with different formation mechanisms. Near the inlet, dro-
results are shown in Fig. 16 on a xy-plane. Due to the large velocity plets are mainly stripped off the interface by the gas stream, as
difference between the liquid and gas jets, a Kelvin–Helmholtz highlighted by the purple arrow in Fig. 16(a)–(c). Further down-
instability develops when the two streams meet at the end of the stream, the droplet formation is mainly due to ligament
separator plate (Marmottant and Villermaux, 2004; Fuster et al., break-up, or a combination of the two mechanisms, see
2013). The interface wave appearing near the inlet, which is indi- Fig. 16(d)–(f).
cated by a purple arrow, is at first characterized by a small ampli- Droplets are not only generated at different streamwise posi-
tude and by a shape close to a sinusoidal wave, see Fig. 16(a). As tions by different mechanisms, but their dynamics changes
the wave propagates downstream, its amplitude grows and the remarkably as well, as shown in Fig. 17 by the trajectories of the
wave crest rises into the fast gas stream, introducing a significant droplets formed near the two streamwise positions x1 ¼ 2 H and
perturbation into the gas stream that strips small droplets away x2 ¼ 4 H, at times t1 ¼ 0:46 s and t 2 ¼ 0:49 s, respectively. In the
from the wave, see Fig. 16(b). Since the wave propagates much figure are also plotted a snapshot of the interface and the stream-
slower than the gas stream, it acts like an obstacle to the gas flow. wise velocity component, on the plane z ¼ 0, to show the influence
As a consequence, the gas flow over the wave separates and forms of the gas-interface interaction on the droplet dynamics.
a recirculation zone, indicated by a yellow arrow in Fig. 16. The In particular in Fig. 17(a) at time t1 , droplets are formed at the
recirculation zone formed at the downstream of the interface wave wave crest when its amplitude is still relatively small and the
has also been observed experimentally and numerically by Jerome interaction with the gas stream is not very important. As a result,
et al. (2013). The interaction between the wave and the gas stream the gas stream remains fairly horizontal and the flying angle of
becomes more complex as the wave further grows and propagates the droplet is generally small (a flying angle equal to 0 is along
downstream, and the wave crest breaks up in a more pronounced the x-direction). Except for a few large droplets that cross the
way and generates a large amount of droplets, as shown in whole gas stream, most of them stay in its interior as they are con-
Fig. 16(b)–(d). The interface then starts to fold and liquid ligaments vected downstream.
are formed, see Fig. 16(e). The gas stream is significantly influenced The flying angle statistics for the droplets in the sampling rect-
by the deformed interface and is observed to bend upward. A coun- angular cuboid A of Fig. 17(a), and defined by the coordinates
terclockwise circulation is formed near the top of the gas stream, ranges 4:5 H 6 x 5:5 H; H 6 y 6 3 H, and 0 6 z 6 H, is shown in
indicated by blue arrows Fig. 16(a), (b) and (e), (f). The clockwise Fig. 18. The time interval for the sampling is 0:45 6 t 6 0:47 s,
circulation at the downstream of the wave grows in time and which is chosen to select the droplets that are formed close to
pushed downstream as the wave moves. These circulation regions the inlet. From Fig. 17(a), it can be observed that the droplet trajec-
are observed to have a significant impact on droplet formation tories are quite straight after the early acceleration within the gas
(Jerome et al., 2013). The ligaments, stretched by the gas stream, stream. Therefore, it is expected that the flying angles vary little
finally break up into many droplets due to a Plateau-Rayleigh while the droplets are moving within the sampling region. It can
Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 137
Fig. 16. Numerical results with the combined model of the atomizing liquid jet for times in the range 0.44–0.49 s, viewed along the z-direction. The gray and orange colors
indicate the interface resolved by VOF and the small droplets represented by LPP model, respectively. The vector plot on the background shows the flow velocity, with its
magnitude indicated by both length and color of the arrows. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of
this article.)
2D-view 3D-view
(a) t=0.46 s
Fig. 17. 2D and 3D views of trajectories of the droplets formed at the top of the perturbation wave on the liquid–gas interface at times: (a) t ¼ 0:46 s, (b) t ¼ 0:49 s. The
interface and streamwise fluid velocity are also shown in the background.
be clearly seen in Fig. 18(a) that most of the droplets are evenly stream, so when they break-up the droplets are ejected into the
distributed with a flying angle ranging from 30 to 30 degrees. bended gas stream. As a result, most of the droplets are ejected
The statistics include all droplets, hence represented by either with a positive angle and the corresponding flying angle distribu-
the VOF method or the LPP model. tion is shifted to the right as shown in Fig. 18(b). The sampling is
On the other hand, in Fig. 17(b) at time t 2 the wave crest is near done in the same volume A, but within the time interval
x2 , the interface has deformed seriously causing the gas stream to 0:48 6 t 6 0:5 s, which mainly captures the droplets formed by
bend upward. The liquid ligaments are mainly aligned with the the ligaments break-up, as shown in Fig. 17(b). Because of that,
138 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
0.5 0.5
t=4.5-4.7ss
t=0.45-0.47 t=4.8-5.0ss
t=0.48-0.50
0.4 0.4
N/sum (N)
N/sum (N)
0.3 0.3
0.2 0.2
0.1 0.1
0 0
-90 -60 -30 0 30 60 90 -90 -60 -30 0 30 60 90
Flying Angle (deg) Flying Angle (deg)
(a) (b)
Fig. 18. Numerical results of droplet flying angle distribution for the two time intervals: (a) 0:45 < t < 0:47 s, (b) 0:48 < t < 0:50 s. The sampling volume is defined by
4:5H < x < 5:5H; H < y < 3H, and 0 < z < H, and is also indicated by the letter A and the dotted yellow line in Fig. 17(a). (For interpretation of the references to colour in this
figure legend, the reader is referred to the web version of this article.)
the droplets are more aligned with the gas stream, their trajecto- Wep < 3, in terms of both mass and particle number. The droplet
ries are less curved than for the droplets of Fig. 18(a) and their fly- deformation is then expected to be small, and the assumption of
ing angle distribution in Fig. 18(b) is more peaked. a spherical shape in the LPP model is indeed rather good.
The flying angle statistics for the droplets in the sampling rect- The LPP droplet Reynolds number is plotted in Fig. 21 as a func-
angular cuboid B of Fig. 17(b), and defined by the coordinates tion of the ratio between the droplet and cell sizes at two different
ranges 4:5 H 6 x 5:5 H; 2:5 H 6 y 6 3:5 H and for the longer time time intervals. The scaling estimate of the relationship between the
interval 0:4 6 t 6 0:5 s, is shown in Fig. 19. The sampling volume Reynolds number and the diameter of solid particles by
B is similar to that of the experiment of Descamps et al. (2008) Balachandar (2009) can be applied to the droplets here. When
and the time period is long enough to cover the overall cycle of the droplet response time sp is larger than the largest ambient flow
wave development and break up, as in Fig. 16. The simulation time scale, the relative velocity is mainly controlled by the ambient
results agree fairly well with the experimental data, as most of flow velocity and can be considered as independent of dp . In this
the droplets are released with a positive flying angle and the peak limit it can be shown that Rep dp . On the other hand, when sp
of the distribution is at about 20 degrees. However, the experimen- is smaller than the smallest ambient flow time scale, such as the
tal distribution profile is narrower than the numerical one. Kolmogorov time scale, then the relative velocity can be estimated
by the Equilibrium Eulerian Approach as ju ~ up j sp Du
~ =Dt, where
Du~ =Dt is the acceleration of the ambient flow and is again indepen-
Droplet Reynolds and Weber number statistics dent of dp (Ferry and Balachandar, 2001). Then it can be easily
When the LPP model is applied, the relative velocity between 3
shown that in this limit, Rep dp . It is observed from Fig. 21(a)
the droplet and the surrounding fluid can be easily obtained, being
the term u ~ up , to compute the corresponding Reynolds and and (b) that, the variation of the droplet Reynolds number with
the droplet diameter lies in between these two limits, for both
Weber numbers. The probability density function (PDF) and
0:45 < t < 0:47 s and 0:48 < t < 0:5 s, although more droplets with
mass-weighted probability density function (mPDF) of the LPP dro-
smaller Reynolds number are formed for the latter time period.
plets as a function of the Reynolds and Weber numbers are shown
Furthermore, the droplets with larger Reynolds number agree bet-
in Fig. 20, for the sampling volume A of Fig. 17(a) and the two time
intervals 0:45 6 t 6 0:47 s and 0:48 6 t 6 0:5 s. In the first time ter with Rep dp ; while the droplets with smaller Reynolds num-
interval about 45% of the total number of droplets is represented ber follow the Rep dp scaling. At last, Fig. 21 also shows that there
by the LPP model, and about 64% in the second one. are a significant amount of droplets with large Reynolds number
In general, the results for the two time intervals are rather sim- and small diameter-to-cell-size ratio formed in atomization. It is
ilar, even if slightly more droplets are observed in the PDF of the expected the dynamics of these droplets would be incorrect if they
second time interval at small Reynolds and Weber numbers. Both remain to be resolved by VOF with low resolution (dp =h K 4). This
PDFs decrease with Rep and while most of the droplets are in a confirms it is necessary to employ LPP model to represent these
regime of very small Reynolds numbers, Rep < 1, still there is a sig- droplets.
nificant amount of droplets in the flow at higher numbers, in the
range 1 < Rep < 200 as shown in Fig. 20(a). If the statistics is made Comparison of droplet trajectories with experiment
on mass rather than droplet number, the mass-weighted PDF is The trajectories of a sample of large droplets generated by the
obtained. From Fig. 20(c) a large fraction of the total mass is in numerical simulation of the atomization process are shown in
the droplets with 20 < Rep < 70. It should be reminded that the Fig. 22(a), while in the experiment the droplet trajectories are
criterion dp < 4Dx is used for RD-to-LPP conversion. For this range traced through high-speed cameras (Descamps et al., 2008). Due
of Reynolds numbers, it is expected that the droplets would not to the limited size of the computational domain, the comparison
have been resolved accurately by VOF, had they remained RD. between numerical and experimental trajectories can be done only
The LPP model is actually a better alternative for these poorly for a few droplets. The traced trajectories are characterized by a
resolved droplets with a relatively large Reynolds number. large positive flying angle, and they are likely to correspond to dro-
Both PDF and mPDF decrease monotonically as a function of the plets with a big inertia. As a matter of fact, from the simulation
Weber number, Wep , and for most droplets it is indeed quite small, results we can compute trajectories of large droplets, with
Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 139
0.5 0.5
Exp Model
t=0.4-0.5 s
0.4 0.4
N/sum (N)
N/sum (N)
0.3 0.3
0.2 0.2
0.1 0.1
0 0
-30 0 30 60 90 -30 0 30 60 90
Flying Angle (deg) Flying Angle (deg)
(a) (b)
Fig. 19. Droplet flying angle distribution: (a) experimental data, (b) numerical results. The sampling region is volume B of Fig. 17(b) with 4:5H < x < 5:5H; 2:5 < y < 3:5H for
the time interval 0:4 < t < 0:5 s. The sampling region is similar to the experimental one, with 4:5H < x < 5:5H and 2:5 < y < 4H (Descamps et al., 2008).
10 0 10 2
t=0.45-0.47s t=0.45-0.47s
t=0.48-0.50s t=0.48-0.50s
10 1
10 -1
10 0
10 -2
PDF
10 -1
10 -3
10 -2
10 -4
10 -3
10 -5 10 -4
0 50 100 150 200 0 0.5 1 1.5 2 2.5 3
Rep Wep
(a) (b)
10 -1 10 1
t=0.45-0.47s t=0.45-0.47s
t=0.48-0.50s t=0.48-0.50s
10 -2 10 0
mPDF
mPDF
10 -3 10 -1
10 -4 10 -2
0 50 100 150 200 0 0.5 1 1.5 2 2.5 3
Re p We p
(c) (d)
Fig. 20. Probability density function (PDF) and mass-weighted probability density function (mPDF) as a function of the Reynolds and Weber numbers, for the droplets in the
sampling volume A of Fig. 17(a), with 4:5H < x < 5:5H; H < y < 3H, and 0 < z < H, in the two time intervals 0:45 < t < 0:47 s and 0:48 < t < 0:5 s.
dp ¼ 0:24—0:5 mm, which match quite well the experimental mea- The corresponding Reynolds and Weber numbers along the dro-
surements. Both the numerical and experimental trajectories are plets trajectory are shown in Fig. 22(b)–(c). For the largest droplet
rather straight. The computed trajectories for the two droplets with diameter dp ¼ 0:49 mm, which belongs to the tail of the PDF
with diameter dp equal to 0.24 and 0.31 mm, curve slightly as of Fig. 20, the Reynolds number can be as high as 400, while the
the droplets enters the faster gas stream and are subject to a sud- Weber number remains smaller than 5. These two numbers vary
den acceleration in the streamwise direction. in a remarkable way as the droplets move downstream, though
140 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
10 2 10 2
Rep~dp3
Re p
Re p
10 1 10 1 Rep~dp3
10 0 10 0
1.00 2.00 4.00 1.00 2.00 4.00
dp/h dp/h
(a) (b)
Fig. 21. Droplet Reynolds number as a function of the ratio between the droplet and cell sizes for the droplets in the sampling volume A of Fig. 17(a), with
4:5H < x < 5:5H; H < y < 3H, and 0 < z < H, in the two time intervals 0:45 < t < 0:47 s and 0:48 < t < 0:5 s.
0.04
0.03
y (m)
0.02 dp
dp=0.24mm
Gas Jet dp
dp=0.28mm
dp
dp=0.31mm
0.01 dp
dp=0.49mm
Liquid Jet Exp
0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
x (m)
(a)
600 6
ddp=0.24mm
p dp
dp=0.24mm
500 ddp=0.28mm
p dp
5 dp=0.28mm
ddp=0.31mm
p dp
dp=0.31mm
400 ddp=0.49mm
p dp
4 dp=0.49mm
Wep
Rep
300 3
200 2
100 1
0 0
0 0.02 0.04 0.06 0.08 0 0.02 0.04 0.06 0.08
x (m) x (m)
(b) (c)
Fig. 22. Numerical results of a sample of large droplets as a function of the longitudinal x-direction: (a) projection of the trajectories on the xy-plane, (b) Reynolds number,
and (c) Weber number. The numerical trajectories are also compared with the experimental measurements.
the trajectories are quite smooth. These fluctuations are related to in Fig. 23(a) and (b), respectively. The comparison of the probabil-
the interaction with the gas turbulent flow. The turbulence of the ity density functions (PDF) of the droplet size for the combined
gas flows is shown in background of Fig. 17. However, the inertia model and DNS on the same size mesh is also shown in Fig. 23(d).
of these droplets is quite large, and the influence of turbulence The results of the combined model for the droplets larger than
on their trajectory is not profound. the grid size agree very well with the DNS result on the same size
mesh. This indicates that the LPP coupling and conversion model
Droplet size distribution do not introduce error on the droplet formation, which is mainly
Finally, the size distribution for the droplets formed in the resolved by VOF. When the LPP model is introduced, a larger num-
gas-assisted atomization is shown in Fig. 23. The droplet size dis- ber of droplets smaller than the cell size are observed, see
tributions computed by the combined model and DNS are shown Fig. 23(a). The number of smaller-than-cell-size droplets is
Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 141
10 3 10 3
counts
counts
10 2 10 2
10 1 10 1
10 0 10 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
d (mm) d (mm)
(a) (b)
10 2
DNS, h=H/128 Model, h=H/64
DNS, h=H/64
10 5 DNS, h=H/128
10 1
10 4
counts
10 0
10 3
10 -1
10 2
10 1 10 -2
0 0.5 1 1.5 2 2.5 0 0.05 0.1 0.15 0.2 0.25
d (mm) d/H
(c) (d)
Fig. 23. Size distributions for droplets formed in atomization. (a) Counts of droplets for present model on the mesh h ¼ H=64 ¼ 0:156 mm, (b) Counts of droplets for DNS on
the mesh h ¼ H=64 ¼ 0:156 mm, (c) Counts of droplets for DNS on the mesh h ¼ H=128 ¼ 0:078 mm, (d) Probability density function (PDF) (only droplets that are larger than
the cell size are considered in PDF calculation.) The characteristic length scales for the original and finer meshes are k1 ¼ 0:166 mm and k2 ¼ 0:095 mm, respectively.
significantly lower for the DNS results without the LPP model as In order to see the effect of grid resolution a DNS simulation is
shown in Fig. 23(b). The discrepancy may be due to the erroneous conducted with a finer mesh, h ¼ H=128. The total number cells is
dynamics of small droplets when they remain to be resolved by about 67 million and it cost about 400 h on 2048 cores to reach
VOF. Indeed small fragments of VOF-represented liquid tend to similar physical time t ¼ 0:6 s. The detailed results of the
be reconstructed and advected so erroneously that they do not finer-mesh DNS are not the focus of the present paper and discus-
move at all, an effect that can be easily demonstrated but is never sion of which will be relegated to future work.
discussed as the dynamics of such small fragments is considered The droplet size distribution obtained by DNS on a finer mesh is
largely erroneous anyway. As these small fragments stay where shown in Fig. 23(c). The comparison of the PDF of the droplet size
they are formed they are easily swept by and merged into the mov- with the two different meshes is shown in Fig. 23(d). With similar
ing bulk liquid jet. If these small fragments are converted to LPP, sampling time period (about 0.03 s), the droplet number for the
then they can be advected away from the liquid jet and remain finer mesh is an order of magnitude larger than that for the original
conserved. Since the size distributions for both the combined mesh. It is also seen that the size distribution of small droplets
model and DNS are possibly erroneous for the size range dp < h. dp < 0:2 mm is slightly better resolved, but not yet converged,
The droplets smaller than the cell size are thus excluded in the when the finer mesh is used. Exponential decay of droplet number
PDF statistics shown in Fig. 23(d). similar to the original mesh is observed for dp 2 ½0:1; 0:6 mm in the
Furthermore, an exponential decay of droplet number with dro- fine mesh results, though the decay is faster. The characteristic
plet diameter for dp 2 ½0:2; 1:1 mm is observed in both Fig. 23(a) length scale for the exponential decay, k2 ¼ 0:095 mm, is 42%
and (b). The decay rate can be characterized by a length scale k1 , smaller than k1 ¼ 0:166 for the original mesh, indicating lack of
which is considered as the ‘‘average ligament size’’ in convergence, so that yet unresolved small scales in the flow are
Marmottant and Villermaux (2004). A least square fit of the data important to droplet formation. To yield an accurate prediction
of the present model for dp 2 ½0:2; 1:1 mm yields k1 ¼ 0:166 mm. of the droplet size distribution, including the expected peak at
The exponential decay with k1 ¼ 0:166 is shown to match very small scales and the exponential decay, the smaller scales must
well with the DNS results on the same size mesh, see Fig. 23(b). be resolved with sufficient grid resolution. It is likely that the com-
It is noted that k1 is comparable to the cell size h ¼ 0:156 mm. putational cost for a fully-resolved DNS would be extremely high.
142 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
Therefore, it may be unavoidable for future exploration of atomiza- acknowledge Grants x20142b6115 and x20142b7325 from
tion in this parameter range to develop subgrid models accounting GENCI. The authors acknowledge Dr. G. Tryggvason and Dr. S.
for the small-scale physics, such as interface instability, Dabiri for communicating their code, which serves as the initial
interface-turbulence interaction and droplet formation at the rele- version of PARIS-Simulator. We also thank T.J. Arrufat, Dr. D.
vant scales. Fuster, L. Malan, and Dr. P. Yecko for their contributions to the
development of PARIS-Simulator.
Conclusions
Hirt, C.W., Nichols, B.D., 1981. Volume of fluid (VOF) method for the dynamics of Mei, R., Adrian, R.J., 1992. Flow past a sphere with an oscillation in the free-stream
free boundaries. J. Comput. Phys. 39, 201–225. velocity and unsteady drag at finite Reynolds number. J. Fluid Mech. 237, 323–
Jerome, J.J.S., Marty, S., Matas, J.-P., Zaleski, S., Hoepffner, J., 2013. Vortices catapult 341.
droplets in atomization. Phys. Fluids 25, 112109. Ménard, T., Tanguy, S., Berlemont, A., 2007. Coupling level set/vof/ghost fluid
Josserand, C., Lemoyne, L., Troeger, R., Zaleski, S., 2005. Droplet impact on a dry methods: validation and application to 3d simulation of the primary break-up
surface: triggering the splash with a small obstacle. J. Fluid Mech. 524, 47–56. of a liquid jet. Int. J. Multiphase Flow 33, 510–524.
Labourasse, E., Lacanette, D., Toutant, A., Lubin, P., Vincent, S., Lebaigue, O., Pope, S.B., 2000. Turbulent Flows. Cambridge University Press.
Caltagirone, J.-P., Sagaut, P., 2007. Towards large eddy simulation of isothermal Popinet, S. The basilisk code. <http://basilisk.fr/>.
two-phase flows: governing equations and a priori tests. Int. J. Multiphase Flow Popinet, S., 2003. Gerris: a tree-based adaptive solver for the incompressible euler
33, 1–39. equations in complex geometries. J. Comput. Phys. 190, 572–600.
Lafaurie, B., Nardone, C., Scardovelli, R., Zaleski, S., Zanetti, G., 1994. Modelling Popinet, S., 2009. An accurate adaptive solver for surface-tension-driven interfacial
merging and fragmentation in multiphase flows with SURFER. J. Comput. Phys. flows. J. Comput. Phys. 228, 5838–5866.
113, 134–147. Renardy, Y., Renardy, M., 2002. PROST: a parabolic reconstruction of surface tension
Lakehal, D., Labois, M., Narayanan, C., 2012. Advances in the large-eddy and for the volume-of-fluid method. J. Comput. Phys. 183, 400–421.
interface simulation (leis) of interfacial multiphase flows in pipes. Prog. Rudman, M., 1997. Volume-tracking methods for interfacial flow calculations. Int. J.
Comput. Fluid Dyn. 12, 153–163. Numer. Methods Fluids 24, 671–691.
Lakehal, D., Liovic, P., 2011. Turbulence structure and interaction with steep Rudman, M., 1998. A volume-tracking method for incompressible multifluid flows
breaking waves. J. Fluid Mech. 674, 522–577. with large density variations. Int. J. Numer. Methods Fluids 28, 357–378.
Larocque, J., Vincent, S., Lacanette, D., Lubin, P., Caltagirone, J.-P., 2010. Parametric Scardovelli, R., Zaleski, S., 1999. Direct numerical simulation of free-surface and
study of LES subgrid terms in a turbulent phase separation flow. Int. J. Heat interfacial flow. Annu. Rev. Fluid Mech. 31, 567–603.
Fluid Flow 31, 536–544. Schmidt, D.P., Rutland, C.J., 2000. A new droplet collision algorithm. J. Comput. Phys.
Lasheras, J.C., Villermaux, E., Hopfinger, E.J., 1998. Break-up and atomization of a 164, 62–80.
round water jet by a high-speed annular air jet. J. Fluid Mech. 357, 351–379. Shinjo, J., Umemura, A., 2010. Simulation of liquid jet primary breakup: dynamics of
Le Chenadec, V., Pitsch, H., 2013. A monotonicity preserving sharp interface flow ligament and droplet formation. Int. J. Multiphase Flow 36, 513–532.
solver for high density ratio two-phase flows. J. Comput. Phys. 249, 185–203. Sirignano, W.A., 2005. Volume averaging for the analysis of turbulent spray flows.
Lebas, R., Menard, T., Beau, P.A., Berlemont, A., Demoulin, F.-X., 2009. Numerical Int. J. Multiphase Flow 31, 675–705.
simulation of primary break-up and atomization: Dns and modelling study. Int. Smagorinsky, J., 1963. General circulation experiments with the primitive
J. Multiphase Flow 35, 247–260. equations. Monthly Weather Rev. 91, 99–164.
Legendre, D., Magnaudet, J., 1997. A note on the lift force on a spherical bubble or Sundaram, S., Collins, L.R., 1996. Numerical considerations in simulating a turbulent
drop in a low-Reynolds-number shear flow. Phys. Fluids 9, 3572–3574. suspension of finite-volume particles. J. Comput. Phys. 124, 337–350.
Leonard, B.P., 1979. A stable and accurate convective modelling procedure based on Sussman, M., Smereka, P., Osher, S., 1994. A level set approach for computing
quadratic upstream interpolation. Comput. Methods Appl. Mech. Eng. 19, 59– solutions to incompressible two-phase flow. J. Comput. Phys. 114, 146–159.
98. Tomar, G., Fuster, D., Zaleski, S., Popinet, S., 2010. Multiscale simulations of primary
Ling, Y., Haselbacher, A., Balachandar, S., 2010. A numerical source of small-scale atomization. Comput. Fluids 39, 1864–1874.
number-density fluctuations in Eulerian–Lagrangian simulations of multiphase Toutant, A., Chandesris, M., Jamet, D., Lebaigue, O., 2009a. Jump conditions for
flows. J. Comput. Phys. 229, 1828–1851. filtered quantities at an under-resolved discontinuous interface. Part 1:
Ling, Y., Parmar, M., Balachandar, S., 2013. A scaling analysis of added-mass and Theoretical development. Int. J. Multiphase Flow 35, 1100–1118.
history forces and their coupling in dispersed multiphase flows. Int. J. Toutant, A., Chandesris, M., Jamet, D., Lebaigue, O., 2009b. Jump conditions for
Multiphase Flow 57, 102–114. filtered quantities at an under-resolved discontinuous interface. Part 2: A priori
Lomholt, S., Maxey, M.R., 2003. Force-coupling method for particulate two-phase tests. Int. J. Multiphase Flow 35, 1119–1129.
flow: Stokes flow. J. Comput. Phys. 184, 381–405. Tryggvason, G., Bunner, B., Esmaeeli, A., Juric, D., Al-Rawahi, N., Tauber, W., Han, J.,
Lomholt, S., Stenum, B., Maxey, M.R., 2002. Experimental verification of the force Nas, S., Jan, Y.J., 2001. A front-tracking method for the computations of
coupling method for particulate flows. Int. J. Multiphase Flow 28, 225–246. multiphase flow. J. Comput. Phys. 169, 708–759.
Marmottant, P., Villermaux, E., 2004. On spray formation. J. Fluid Mech. 498, 73– Tryggvason, G., Scardovelli, R., Zaleski, S., 2011. Direct Numerical Simulations of
111. Gas–Liquid Multiphase Flows. Cambridge University Press.
Matas, J.-P., Marty, S., Cartellier, A., 2011. Experimental and analytical study of the Vincent, S., Balmigère, G., Caltagirone, J., Meillot, E., 2010. Eulerian–Lagrangian
shear instability of a gas–liquid mixing layer. Phys. Fluids 23, 094112. multiscale methods for solving scalar equations-application to incompressible
Maxey, M.R., Patel, B.K., 2001. Localized force representations for particles two-phase flows. J. Comput. Phys. 229, 73–106.
sedimenting in Stokes flow. Int. J. Multiphase Flow 27, 1603–1626. Vincent, S., Brändle de Motta, J.C., Sarthou, A., Estivalezes, J.-L., Simonin, O., Climent,
Maxey, M.R., Patel, B.K., Chang, E.J., Wang, L.P., 1997. Simulations of dispersed E., 2014. A Lagrangian VOF tensorial penalty method for the DNS of resolved
turbulent multiphase flow. Fluid Dyn. Res. 20, 143–156. particle-laden flows. J. Comput. Phys. 256, 582–614.
Maxey, M.R., Riley, J.J., 1983. Equation of motion for a small rigid sphere in a
nonuniform flow. Phys. Fluids 26, 883–889.